ArtBoost: Synthetic Articulatory Data Augmentation for Acoustic-to-Articulatory Inversion
Researchers have developed ArtBoost, a new data augmentation technique to improve acoustic-to-articulatory inversion (AAI) models. This method utilizes large-scale speech-mesh datasets, originally created for 3D facial animation, to generate pseudo articulatory trajectories. These synthetic trajectories are used for pre-training AAI models before fine-tuning with limited real electromagnetic articulography (EMA) data, leading to consistent performance gains in metrics like PCC and RMSE. AI
IMPACT Enhances AI's ability to model speech articulation, potentially improving speech synthesis and recognition systems.